{"paper":{"title":"Understanding Generalization through Decision Pattern Shift","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Decision pattern shifts in neural networks correlate linearly with generalization gaps, framing failure as internal mechanism drift.","cross_cats":["cs.CV"],"primary_cat":"cs.LG","authors_text":"Hongbin Pei, Huiqi Deng, Peng Zhang, Quanshi Zhang, Xia Hu, Yibo Li","submitted_at":"2026-05-13T08:14:59Z","abstract_excerpt":"Understanding why deep neural networks (DNNs) fail to generalize to unseen samples remains a long-standing challenge. Existing studies mainly examine changes in externally observable factors such as data, representations, or outputs, yet offer limited insight into how a model's internal decision mechanism evolves from training to test. To address this gap, we introduce Decision Pattern Shift (DPS), a new perspective that defines generalization through the stability of internal decision patterns and quantifies failure as their deviation from those learned during training. Specifically, we repre"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"the DPS magnitude correlates linearly with the generalization gap (nearly all Pearson r > 0.8), revealing generalization as a systematic drift in the model's internal decision mechanism","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that a GradCAM-based channel-contribution vector faithfully represents the model's internal decision pattern and that deviation from the class-average pattern is the correct way to quantify harmful shift","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DPS quantifies deviation of per-sample decision patterns from class averages and shows linear correlation with generalization gaps while unifying degradation scenarios into a continuous trajectory.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Decision pattern shifts in neural networks correlate linearly with generalization gaps, framing failure as internal mechanism drift.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"663f39aaeea4662c35aa965ba24d8aa299b5cc0a358ff605dd5f0df99750565f"},"source":{"id":"2605.13148","kind":"arxiv","version":1},"verdict":{"id":"3f14911c-cf0e-4f9b-82e4-9b057621cf5d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:15:47.440458Z","strongest_claim":"the DPS magnitude correlates linearly with the generalization gap (nearly all Pearson r > 0.8), revealing generalization as a systematic drift in the model's internal decision mechanism","one_line_summary":"DPS quantifies deviation of per-sample decision patterns from class averages and shows linear correlation with generalization gaps while unifying degradation scenarios into a continuous trajectory.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that a GradCAM-based channel-contribution vector faithfully represents the model's internal decision pattern and that deviation from the class-average pattern is the correct way to quantify harmful shift","pith_extraction_headline":"Decision pattern shifts in neural networks correlate linearly with generalization gaps, framing failure as internal mechanism drift."},"references":{"count":82,"sample":[{"doi":"","year":2017,"title":"A closer look at memorization in deep networks","work_id":"1b682202-7721-4e0a-b99c-ea2b70420c42","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Nearly-tight vc-dimension and pseudodimension bounds for piecewise linear neural networks","work_id":"c11d7346-0766-4a72-9137-10212a7cf800","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Network dissection: Quantifying interpretability of deep visual representations","work_id":"81485029-2003-48ff-82fa-04c054e4fcb9","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2006,"title":"Analysis of representations for domain adaptation","work_id":"b8163d18-21f0-490f-a1b2-ccd82a5a918d","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Estimating generalization under distribution shifts via domain-invariant representations","work_id":"00a5c439-e7a1-4482-b7ca-1c280d3fba6e","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":82,"snapshot_sha256":"7244bae23fb04d44a3e223464163da28caeb3c2bd6c4becbce6960363ffd4564","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ff424f1b4f8646379b5371f6251ce05da1011f653874e4c3c06b74e7f5f721ae"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}